AI agents for business: what actually works

A lead came in at half past one in the morning. By nine, when the rep opened their laptop, the customer had already answered four questions, named a budget, and received a link to sample work. All the rep had left to do was call and hash out the details. No magic — this is an AI agent I built and shipped in two weeks.

I've collected enough stories like that over the past year to fill an article. And roughly as many where an agent wasn't needed at all — and I talked people out of it. We'll cover both lists: no "digital employees," no "revolution," none of the marketing noise this topic is currently buried under.

To be clear in one line: an AI agent is a neural network (an LLM) that has access to your systems through an API and a set of boundaries within which it decides for itself what to do: look something up in the CRM, ask a question, record the result, call in a human. Anything that responds by a rigid script isn't an agent but a chatbot, whatever the slide deck calls it.

Tasks where AI agents pay off

The most rewarding one is qualifying inbound leads. The agent replies to a lead first, pulls the essentials out of the customer (what they need, when, what budget), filters out spam, and hands the live ones to a rep — with a ready summary. At night, on weekends, in seconds. This is the scenario that's been running the longest for me, and there's a separate breakdown of it with the prompt and the rollout mistakes: AI agent for lead qualification.

The second is dull but a real time-eater: sorting the inbox. Email, Telegram, forms — the agent splits the flow into leads, order questions, documents, and junk, and each type goes to the right person. Nobody wants to pay for a "mail sorter" until they add up how many hours a week the team loses to it.

The third is content and monitoring: the agent watches sources, picks out what's relevant, and prepares draft posts and summaries. The human stays the editor. A couple of setups like this, with the architecture, are in the piece on marketing automation with n8n.

And the fourth, my favorite, is an AI analyst on top of the business's data. You ask in Telegram in plain language: "how much did we spend on ads in June and what did it bring in?" — and half a minute later you get an answer in numbers pulled from your own sales and ad accounts. Not a report from a marketer by Friday, but an answer now. There'll be a separate breakdown of this setup soon — together with the dashboard it sits on.

Where you shouldn't call in an agent

Twice in the past year I talked clients out of a rollout. Not because I couldn't build it — I could have, and taken the money. Because it wouldn't have paid off, and three months later AI agents at that company would be written off as "nonsense, we tried it."

The first case — they wanted a "salesperson agent" that closes deals on its own. Average deal size close to a million, a sales cycle of months, tons of negotiation and haggling. In a sale like that an agent can do one thing: quickly and cleanly prepare the ground for a human. Hand it the whole deal and you burn a lead that cost a lot to get.

The second case is subtler, and it's also the most common. A company asks for an agent while half the deals in its own CRM have no stages, UTM tags aren't set, and there's no standard for responding to a lead — every rep answers however they can. An agent won't create order; it demands it. You can automate chaos, and you'll get automated chaos — just faster and more expensive.

And separately: anywhere a mistake is costly — money, legal, health — the agent prepares a draft for a human but doesn't act on its own. That's not a limitation of the technology, it's common sense.

What a durable agent is built from

The difference between a demo and a working system isn't the neural network. The model, in fact, is chosen quickly: a cheap one is fine for sorting mail, the analyst needs something smarter. The difference is in the wiring that a demo never shows.

A working agent has it spelled out which actions it's allowed and which it isn't. It's spelled out when it must go quiet and hand the conversation to a human: the customer is angry, the question is unusual, the amount is above a threshold. Every action it takes is written to logs, and for the first few weeks a live person reads those logs — that's how the agent "learns on the job" from real conversations rather than fantasies about them. The agent has no dangerous permissions at all: the worst it can break is handing a rep one extra conversation.

I build this on n8n — it lives on its own server, gets along with any API, and all of a company's agent scenarios sit in one place instead of across five subscriptions.

What an AI agent costs

Cheaper than people assume — and not where they assume. The server is a few dollars a month. The neural network itself is billed by tokens: a typical qualification conversation is cents. Even at hundreds of leads a month, the agent's "life" runs to tens of dollars.

What you pay for in this story is something else — the rollout. Mapping the process, designing the logic, connecting the services, running it against real conversations, tuning it. A simple agent is one to two weeks, a complex one a month. After that it's arithmetic: a rep who used to spend two hours a day on first-touch and sorting gets those hours back. Forty hours a month, every month, no vacations.

How to tell you're ready

Before we start I ask a client three questions. Three "yes"es and the agent will fly.

Is there a task the team does by hand every day by the same logic? Not "lots of routine in general," but a specific one: we answer leads, we sort mail, we assemble a report. An agent is a narrow specialist; "sort out everything" is not something it can do.

Is the process described at least at the level of "what goes in, what comes out, who's responsible"? If the process lives in one employee's head — first we get it out of the head, then we automate.

Is there data the agent can be connected to, and is it in order? A CRM with stages, a knowledge base, spreadsheets. As for what happens when there's no data — see the section on chaos above.

Frequently asked questions

Will an AI agent replace an employee?

No. The agent takes the routine part: first replies, sorting, data gathering, drafts. Decisions and accountability stay with people. The working framing isn't "replace the rep" but "give them back 2-3 hours a day."

How much does it cost to deploy an AI agent?

Running the agent is pennies: a server plus the neural-network fee, cents per conversation. The real money is the rollout: mapping the process, configuration, testing against real conversations. A simple agent launches in 1-2 weeks.

Is it safe to give an agent customer data?

It depends on the architecture. The agent gets the minimum access it needs, personal data can be masked before it's sent to the neural network, and every action is logged. It's a matter of configuration, not a property of the technology.

What happens if the agent makes a mistake?

In a working setup an agent's mistake looks like "handed the conversation to a rep," not "deleted the database": the list of allowed actions, escalation rules, and logs are all set up before launch, and the agent has no dangerous permissions.